feat(application): integrate measured-load ingestion training and planner source

services/netplan-v4/commissioning/APPLICATION_STATUS.md

@@ -0,0 +1,80 @@
+# Integrated application data/model path - 2026-10-02
+
+User request: finish application implementation, stop asking for repeated hardware
+confirmations and adding standalone diagnostic samplers. No actuator permission is
+implied by that development request. Confirmed EV allocation remains 161.44 kWh /
+39 kW each direction, with SDL reserve already excluded.
+
+## Implemented in the existing application
+
+- Native ManagerNetzfahrplanV4DatenTrait: 30-second acquisition of configured numeric
+  sources; durable private app-owned outbox; original timestamps, no new Modbus
+  polls; acknowledged per-plant batches with retries/backoff. Cursor advances only
+  after exact dataset/acceptedThrough acknowledgement. Batches bounded below the
+  existing portal's 1 MiB body limit. No credentials in data payload or reports.
+- Existing device proxy now forwards /planner-v4/measurements through the existing
+  authentication, plant binding, licence check and separate V4 rate limiter.
+  Device route cannot modify the dataset mapping or issue a trial grant.
+- Existing V4 database stores immutable versioned source mappings and observations.
+  Same-time conflicts reject the whole batch; repeated identical data is idempotent.
+  No writes to portal/users/legacy measurement databases.
+- Existing V4 worker aggregates physical load using original source times. Virtual
+  EV/SDL filter outputs are not load inputs. Solar terminal variant counts signed
+  terminal power once. Small unobserved parts remain quantified, never zero-filled.
+  This configured formula remains an estimate, not independent AC/DC certification.
+- Application profile training actually runs daily/weekly. Models are immutable,
+  bootstrap/holdout status explicit. Incumbent replacement requires a causal
+  holdout when available. Existing data and model versions are not relabelled.
+- Three corrected load variants (daily robust profile, last-day profile,
+  weekday/weekend profile) are paired with the existing PV families 3/13/23.
+  They are tagged physical-profile-v1; they are NOT claimed to be the old trained
+  load algorithms under a new data name. Automatic future family registry work
+  and economic replay are still distinct from this load-profile implementation.
+- Selecting corrected_profile in the existing GUI/setting now actually changes
+  the load entering the existing optimizer. Missing trained data cannot silently
+  fall back to the contaminated legacy household series.
+- A current SDL request is an explicitly labelled PERSISTENCE SCENARIO for the
+  shadow economic calculation, not a published future schedule. Missing/stale SDL
+  is not zero. This is not a robust/full SDL production-dispatch policy.
+- GUI now exposes dataset/source selection, records, usable equivalent hours,
+  model version, validation and real training cadence. Existing normal portal
+  app.js/index files are not changed.
+
+## Deployment package
+
+Server: commissioning/deploy_application.py --plant <approved UUID> --apply
+Rebuild/test V4 with Python 3.11, portal routing tests, consistent SQLite backup,
+replace ONLY V4 and portal, configure the new per-plant dataset. No automatic
+forecast-source switch and no control grant. Atomic file-bind replacement requires
+portal recreation, not a simple restart. Additive DB rollback retains new data.
+Default invocation without --apply validates source manifests only.
+
+Test host: /srv/agent/netplan-v4-application-build/install.php inside Symcon.
+Installs a DATA-ONLY patch of the currently installed passive manager (not the
+unreleased controller-trial candidate), reloads EMS, configures acquisition and
+imports up to 48h of the existing 23-source observer as a delivery backlog.
+Original observer and old raw journals remain intact. Existing separate observers
+are NOT automatically stopped by this initial installer; consolidation should
+follow confirmed native batch acceptance. No new root diagnostic category.
+An asynchronous module registration may require one repeat, handled explicitly.
+
+Dataset lihrenmoos-physical-v1: minimum coverage 95%, longest unsupported portion
+10s, at least 24 equivalent usable hours, 28-day profile history. The 95% rule is
+an explicit modelling policy, not a statement that missing energy was measured.
+Insufficient data remains collecting; no promised time to good forecasts.
+All original acquisition thresholds remain unchanged.
+
+## Tests and honest limits
+
+Host tests use existing isolated QA dependencies; native tests mock IPS/HTTP and
+use temporary files. Real container and kernel tests occur on the user-run deploy
+commands, not during preparation. No services/modules/settings/devices were modified
+by this development session.
+
+Still NOT a complete production commissioning: the corrected profile-to-local
+feedback/dispatch integration, full automatic cost replay, independent long-term
+outbox/server retention and staged live/failure acceptance remain open. The trial
+code from previous commits remains gated and is not included in the data-only
+native installation. No accountingEvidenceId/device watchdog proof is fabricated.
+This release is an integrated application build, NOT a claim that the full original
+multi-plant production scope is done. No main/beta release is authorized by tests alone.